---
id: 20261119-T0-14
title: "VisHead-Spec：按注意力头路由视觉信息，提升多模态推测解码速度"
title_en: "VisHead-Spec Routes Visual Attention by Head for Faster Multimodal Speculative Decoding"
url: https://ai.daily.yangsir.net/daily/20261119-T0-14
issue_date: 2026-11-19
publish_date: 2026-11-18T05:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2611.21839
---

# VisHead-Spec：按注意力头路由视觉信息，提升多模态推测解码速度

论文提出VisHead-Spec，针对多模态推测解码在高并发场景下加速效果衰减的问题。推测解码能在保持目标模型输出不变的前提下加速视觉语言模型，但随着批大小增加，系统变为计算受限，加速比下降。该方法通过头级别路由视觉注意力来改善这一状况。对部署VLM推理服务的团队，这关系到高并发下的实际吞吐和成本。

## English Version

**VisHead-Spec Routes Visual Attention by Head for Faster Multimodal Speculative Decoding**

The paper presents VisHead-Spec to address the diminishing speedup of multimodal speculative decoding under high-concurrency serving. Speculative decoding accelerates vision-language models while preserving target-model outputs, but as batch size grows the system becomes compute-bound and gains shrink. The method improves this via head-routed visual attention. For teams deploying VLM inference services, this bears directly on throughput and cost under load.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2611.21839)

**详情页**：https://ai.daily.yangsir.net/daily/20261119-T0-14

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*